Justifying model complexity: Evaluating transfer learning against classical models for intraoperative nociception monitoring under anesthesia

Background Accurate intraoperative detection of nociceptive events is essential for optimizing analgesic administration and improving postoperative outcomes. Although deep learning approaches promise improved modeling of complex physiologic dynamics, their added computational and operational complexity may not translate into clinically meaningful benefit, particularly in small, high-resolution perioperative datasets. Methods We performed a head-to-head evaluation of classical supervised models (L1-regularized logistic regression and 50-, 200-tree Random Forests, with and without drug dosing features) against a Temporal Convolutional Network (TCN) transfer-learning framework for intraoperative nociception detection. Using 101 adult surgical cases with 30 physiologic and 18 drug dosing features sampled in 5-second windows, models were assessed under leave-one-surgery-out cross-validation using AUROC and AUPRC. We further examined probability calibration, multiple ensemble strategies, permutation importance features, and computational cost in terms of inference operations and memory footprint. Results Drug-aware Random Forests of various trees (50 trees vs. 200 trees) achieved the highest discrimination (AUROC 0.716; AUPRC 0.399), outperforming the TCN transfer-learning model (AUROC 0.649; AUPRC 0.311). However, increasing personalization windows in the TCN yielded inconsistent and modest gains (p > 0.05). Isotonic calibration substantially improved probability calibration but did not affect discrimination. No ensemble method surpassed the standalone Random Forest; the gated network consistently assigned >84% weight to the classical model. Computational analysis revealed that while the TCN was more compact in total memory footprint, the smaller, 50-tree Random Forest inference required two orders of magnitude fewer operations, with faster training and lower operational complexity. Conclusions In this clinically realistic benchmark, interpretable classical models operating on well-engineered features without personalization matched or exceeded the performance of a personalized deep learning approach while remaining computationally cheaper and simpler to deploy. These findings underscore the importance of rigorously justifying model complexity in perioperative machine learning and suggest that, for intraoperative nociception monitoring, classical approaches may offer a more favorable balance of accuracy, interpretability, and operational efficiency.